cudadl/controller_main.cc

#include <NTL/ZZ.h>

#include <iostream>

#include <string.h>
#include <stdlib.h>

#include "controller.h"

NTL_CLIENT

// The total number of nodes (that is, dpnodes)
static unsigned short total_nodes = 1;
// The amount of memory we can use per dpnode
static unsigned short GB_mem_per_node = 1;

void desired_resources(const ZZ &order, unsigned short &desired_dpnodes,
    unsigned int &max_workers, unsigned int &dpfreq)
{
    // One point in how many is a DP by default?
    unsigned int dpscale = 1000;
    ZZ sorder = SqrRoot(order);

    // How many DPnodes should we use for a problem of this size?

    // 338 is bytes per DP in the table.  10 is a safety factor.
    ZZ dpnumerator = sorder * 338 * 10;
    ZZ dpdenominator;
    dpdenominator = GB_mem_per_node;
    dpdenominator *= 1000000000UL;  // Convert the above line to GB
    ZZ dpnodes = (dpnumerator / (dpdenominator * dpscale)) + 1;
    if (dpnodes > total_nodes) {
	desired_dpnodes = total_nodes;
	ZZ zzdpscale = dpnumerator / (dpdenominator * total_nodes);
	if (NumBits(zzdpscale) > 31) {
	    dpscale = 4294967295U;
	} else {
	    dpscale = trunc_long(zzdpscale, 31);
	}
    } else {
	desired_dpnodes = trunc_long(dpnodes, 31);
    }

    // How many workers would we like to use?
    ZZ sorder23 = sorder >> 23;
    if (NumBits(sorder23) > 30) {
	// Just use all the workers we can find
	max_workers = 4294967295U;  // 2^32 - 1
    } else {
	max_workers = trunc_long(sorder23,31) + 1;
    }

    // By default, 1 in dpscale points are distinguished points.
    dpfreq = 4294967295U / dpscale;

    // Orders smaller than 100*scale^2 behave specially, in order to
    // avoid DP-free cycles
    ZZ orderlimit;
    orderlimit = 100;
    orderlimit *= dpscale;
    orderlimit *= dpscale;

    if (order < 1000) {
	// Just make every point a DP
	dpfreq = 4294967295U;
    } else if (order < orderlimit) {
	// The frequency of DPs should be 10/sqrt(order) to avoid
	// a DP-free cycle, so dpfreq = (10*2^32)/sqrt(order)
	ZZ f = (to_ZZ(10) << 32) / SqrRoot(order);
	dpfreq = trunc_long(f, 31);
    }
}

static void boundcb(const char *boundaddr, unsigned short boundport)
{
    cout << "Listening on " << boundaddr << ":" << boundport << "\n";
    const char *fdenv = getenv("CONTROLLER_BOUNDCB_FD");
    if (fdenv) {
	int fd = atoi(fdenv);
	if (fd > 2) {
	    write(fd, &boundport, 2);
	    write(fd, boundaddr, strlen(boundaddr));
	    close(fd);
	}
    }
}

#ifdef TEST_DESIRED_RESOURCES
int main(int argc, char **argv)
{
    if (argc != 3) {
	std::cerr << "Usage: " << argv[0] << " num_nodes GB_mem_per_node\n";
	return 1;
    }

    total_nodes = strtoul(argv[1], NULL, 10);
    GB_mem_per_node = strtoul(argv[2], NULL, 10);
    cout << "# log_2(order) dpnodes workers dpfreq\n";
    for (int i=44; i<=92; ++i) {
	unsigned short desired_dpnodes;
	unsigned int max_workers;
	unsigned int dpfreq;
	ZZ order;
	order = 1;
	order <<= i;
	order += 1;
	desired_resources(order, desired_dpnodes, max_workers, dpfreq);
	cout << i << " " << desired_dpnodes << " " << max_workers << " "
		<< dpfreq << "\n";
    }
}
#else
int main(int argc, char **argv)
{
    unsigned short bindport = 0;
    Worklist worklist;

    if (controller_parse_args(argc, argv, bindport, worklist,
				total_nodes, GB_mem_per_node)) {
	std::cerr << "Usage: " << argv[0] << " [-p listenport] [-n num_nodes] [-m GB_mem_per_node] [-r reps] N1 iter1 N2 iter2 ...\n";
	return 1;
    }

    return controller_main(worklist, bindport, boundcb);
}
#endif